Apple Fruitlet Abscission Prediction. I. Development and Evaluation of Reflectance Spectroscopy Models

Abstract

Chemical thinning, the most common and cost-effective thinning method, is conducted during early apple fruit development over a 3- to 4-week period using multiple applications of plant growth regulators. It is critical to provide apple growers with tools to assess the efficacy of chemical thinners quickly and accurately because visible responses are not apparent for up to 2 weeks after application. The objective of this study was to build a model to predict apple fruitlet abscission following a chemical thinner application with in situ reflectance data obtained with a portable visible and near infrared (Vis/NIR) spectrophotometer. Developed models were compared with the currently available fruitlet growth model (FGM). ‘Honeycrisp’ fruitlet diameter and reflectance were measured on dates around a chemical thinner application across a 2-year period. After June drop, measured fruitlets were determined to have either persisted or abscised. Random forest, partial least squares regression, and XGBoost classification models were used to predict fruitlet abscission from reflectance data. Each classification model was developed with 2021, 2022, and combined 2021 + 2022 data. For each dataset, 5-fold cross validation was used to assess three model performance metrics: 1) overall accuracy, 2) recall, and 3) specificity. Datasets tested were either unbalanced, majority class down-sampled, or minority class up-sampled with synthetic minority oversampling technique. In both years, the FGM reliably estimated chemical thinner efficacy 9 days after application. Before this time point, the FGM had low prediction accuracy of the minority class in both years—persisting fruitlets in 2021 and abscising fruitlets in 2022. For reflectance spectroscopy, the developed random forest models that were balanced with synthetic minority over-sampling technique were found to be the best combination in predicting chemical thinner efficacy. The combined 2021 + 2022 dataset overall model accuracy ranged from 84% the day before to 93% at 9 days after thinner application. These results show that Vis/NIR is a promising tool to predict chemical thinner efficacy. This technology had high prediction accuracies over a range of fruitlet abscission potential and two growing seasons. Further development and testing of the model over cultivars, chemical thinner timings, and growing regions would facilitate commercialization of the technology.

Keywords

• crop load management
• pomology